Smartphone-based offline AI for multi-disease retinal screening: Real-world accuracy
Aditya Kelkar, Jai Kelkar, Yash Garg, Harsh H. Jain, Sabyasachi SenguptaObjective
To evaluate the diagnostic accuracy of a multi-disease offline artificial intelligence system (Medios-AI, MAI), integrated into a smartphone-based fundus camera, for simultaneous screening of diabetic retinopathy (DR), glaucoma, and age-related macular degeneration (AMD) in a real-world setting.
Methods
In this prospective cross-sectional study, 193 adults (371 eyes) aged ≥18 years with DR, glaucoma, AMD, or normal fundus were enrolled between May and December 2024. Dilated fundus imaging was performed using the Remidio Fundus on Phone (FoP) and Zeiss Clarus 500 cameras. Ungradable images were excluded. The offline MAI algorithm generated disease-specific reports, which were compared to masked grading of Clarus images by two fellowship-trained ophthalmologists. In ambiguous cases, the AI report defaulted to “either DR or AMD.”
Results
MAI achieved sensitivity of 99.3% (95% CI: 96–100), specificity of 95.7% (95% CI: 92–98), and AUROC of 0.99 for detecting any retinal disease. For glaucoma (n = 109), sensitivity was 98.2% (95% CI: 94–100), specificity 99.0% (95% CI: 97–100), AUROC 0.99. For AMD (n = 56), sensitivity was 88.9% (95% CI: 77–96), specificity 97.5% (95% CI: 95–99), AUROC 0.93. For DR (n = 78), sensitivity was 84.6% (95% CI: 75–92), specificity 99.0% (95% CI: 97–100), AUROC 0.92. Agreement on vertical cup-to-disc ratio between AI and graders ranged from −0.1 to +0.1, with intergrader ICC of 0.97 (P < 0.001 for all comparisons).
Conclusions
MAI demonstrated significant diagnostic accuracy for DR, glaucoma, and AMD using an offline, smartphone-based platform, supporting scalable, point-of-care retinal screening in resource-limited settings.